Prediction of Punching Shear Strength using Metaheuristic Approach of Optimization

Advances in Engineering: An International Journal (ADEIJ), Vol.3, No.4
1
PREDICTION OF PUNCHING SHEAR
STRENGTH USING METAHEURISTIC
APPROACH OF OPTIMIZATION
Gaurav Sarkar1
, Suhail Ahmad Magray 1
and Tanushree Sarkar2
1
Department of Computer Engineering, Dsce, Karnataka, India
2
Department of Biotechnology, GGV,Bilaspur,Chhattisgarh, India
ABSTRACT
The relationship between technology and application of civil engineering is not a new concept.
Over the years civil engineering has encountered a slew of issues, most of them have been
solved with the aid of technology. Prediction of punching shear strength is one such problem
statement which could be solved using a metaheuristic approach of optimization. Numerous
experiments on the punching shear resistance of reinforced concrete slabs have been conducted
by researchers, with positive findings. The actual service life will be shortened due to steel bars’
propensity for corrosion. The main goals of all organizations are to make civil engineering
applications more valuable intrinsically so that people can use them to construct faster so that
resources are used more effectively, and to ultimately improve people’s lives. Using
evolutionary artificial neural networks, internal flat slabs of reinforced concrete can be
predicted for their punching shear strength. It is a hybrid model of an artificial neural network
(ANN) and a Genetic algorithm, a metaheuristic based on natural selection that is a subset of
the larger category of evolutionary algorithms (EA). The experimental findings from 519 flat
slabs tested by various authors starting in 1938 were used in this research. The model tries to
predict the dependent feature, Punching shear resistance, using independent features such as
Shape of the column cross section, Column side or smaller side, Larger side of the column,
Average effective depth in X and Y directions, Average reinforcement ratio in X and Y
directions, Column effective width, Effective width / Effective depth, Concrete compressive
strength, and Steel yield strength. Sometimes, signals are altered at the receiving synapses, and
the processing element adds the weighted inputs. Input from one neuron is sent to another (or
output is sent to the outside world) if it reaches the threshold, and the cycle continues. The
algorithm builds the subsequent population at each stage using members of the current
generation. By using Selection, Crossover and Mutation, we can obtain a set of optimal
parameters that aid in producing effective results. We also contrasted the accuracy attained
using GA with other popularly employed optimizer types like SGD, ADAM and RMSProp. We
have also made use of the benefits of the GA algorithm, such as its adaptability, understanding
ability, and lack of computational complexity.
KEYWORDS
Artificial Neural Network, Reinforced Concrete Flat Slab, Punching Shear Strength, Genetic
Algorithm & Metaheuristic.
1. INTRODUCTION
The world has been progressing on every front and civil engineering has a crucial role to play in
it. There are slew of processes which are used as lethal application in the construction domain
such as: Reinforced concrete slabs[1] are one of the common horizontal load-carrying members
in civil engineering, and widely applied in bridges, ports and hydro-structures. Since there is no
beam in the flat slab under longitudinal load, the punching failure of reinforced concrete slab

2
occurred easily. Many researchers have carried out numerous experiments on the punching shear
resistance[2] of reinforced concrete slabs, and obtained successful results. However, steel bars[3]
are prone to corrosion, which will result in the shortening of the actual service life. In recent
years, with application of technology across all domains, the durability of structure is one of the
prime needs of people. For coastal areas and the areas which consist of usage of chlorides such as
dicing salt, the actual service life of structures is often much lower than their design service life,
resulting in massive losses of resources. Fiber reinforced polymer (FRP) [4] is a material which
has many advantages such as light, high strength and corrosion resistance. In a corrosion
environment, to solve the problem of short actual service life of structure, FRP bars [5] can be
applied as an alternative to steel bars in concrete structures.
Regarding theoretical models, the majority of the computational formulas for punching shear
strength of FRP reinforced concrete slabs were obtained from conventional reinforced
concrete flat and modified to take FRP into account.
ACI 318-14 and GB 50010-2010 are two current design requirements that use the eccentric shear
stress model [6] as its theoretical foundation. A number of mitigation measures have been taken
by organizations working across the application of civil engineering such as Large, industrial
constructions, parking garages, warehouses, high-rise buildings, and hostels are the main
applications for flat slabs [7] . They are employed in situations when beamers are not necessary
or in structures with less framework that don’t require beamers.
The major objective of all the organizations are to increase the intrinsic value of the civil
engineering applications so that people can use it, reduce the time in terms of building it for the
purpose of making use of the resources efficiently and eventually creating an impact on the lives
of people. However, during theoretical derivations, the aforementioned empirical models
incorporated some simplifications; as a result, the empirical models were unable to take into
account all of the significant aspects. Furthermore, typical regression analyses using data from
experiments were used to establish the parameters in the aforementioned empirical models. As a
result, the choice of theoretical models and the calibre of the databases have a significant impact
on the models’ correctness. Some algorithms containing data at their core have surfaced recently
with the advancement of artificial intelligence [8]. Among these algorithms, machine learning has
attracted the most research attention [9].
Failure to punching shear strength is due to a strong localized impact, which results in reinforced
flat slabs and foundations collapse. This catastrophic collapse generally happens at the borders of
the columns. Significant cracks occur during failure [10] . Due to its catastrophic character, it
must be prevented; making the determination of the slab’s punching shear strength important
[11].
In this paper, ANN [12] has been used for the purpose of learning the patterns in the given data
efficiently, assisted by loss function and optimizer. In order to further improve the solution,
genetic algorithm had been used as an optimizer which in turn is reducing the amount of time for
building applications of civil engineering, accuracy of development is being improved.
2. RELATED WORK
Sujith Mangalathua ,Hanbyeol Shin, EunsooChoic, Jong- Su Jeon, titled , “Explainable machine
learning models for punching shear strength estimation of flat slabs without transverse
reinforcement” [13].

3
It explains the significance and contribution of the components that affect the punching shear
strength in the extreme gradient boosting model using the SHapley Additive Explanation
approach. To find the best prediction model for the punching shear strength of flat slabs, this
study takes into account seven machine learning techniques in addition to linear regression,
including ridge regression, support vector regression, decision trees, K-nearest neighbours,
random forests, adaptive boosting, and extreme gradient boosting. The associated coefficient of
variation for the extreme gradient boosting model is 0.09, while the model’s coefficient of
determination is 0.98.
Yuanxie Shen, Linfeng Wu, Shixue Liang paper, titled, “Explainable machine learning-based
model for failure mode identification of RC flat slabs without transverse reinforcement”
explained that for determining the failure mechanism of flat slabs, an accurate prediction model is
built by screening 8 machine learning-based models (LR, ANN, DT, SVC, RF, AdaBoost,
GBDT, XGBoost). [14] SHAP provides an explanation for the XGBoost prediction, with the
findings encompassing both general and specific interpretations as well as the feature dependency
relationship between input variables. The best model is XG- Boost, whose precision; recall, F1
score, and accuracy are, respectively, 97.30.
Shasha Lu, Mohammadreza Koopialipoor , Panagiotis G. Asteris , Maziyar Bahri ,Danial Jahed
Armaghani paper, titled, “A Novel Feature Selection Approach Based on Tree Models for
Evaluating the Punching Shear Capacity of Steel Fiber-Reinforced Concrete Flat Slabs”.
This work uses tree predictive models, including random forest (RF), random tree (RT), and
classification and regression trees, to create a new model that can predict the punching shear
capacity of SFRC flat slabs (CART). It also made use of a cutting-edge feature selection (FS)
method. The experiments’ findings showed that the FS-RT model performed better in terms of
prediction accuracy than the FS-RF and FS-CART models. According to measurements of R2
and RMSE, which ranged from 0.9476 to 0.9831 and 14.4965 to 24.9310, respectively, the FS-
RT hybrid approach performed the best in this regard. The three hybrid approaches presented in
this work, FS-RT, FS-RF, and FS-CART, were found to be applicable for forecasting SFRC flat
slabs.[15]
Duy-Thang Vu , Nhat-Duc Hoang, paper, titled, “Punching shear capacity estimation of FRP-
reinforced concrete slabs using a hybrid machine learning approach” explained To develop a new
model that can forecast the punching shear capacity of SFRC flat slabs, this work used tree
predictive models, including random forest (RF), random tree (RT), and classification and
regression trees (CART). It also used a novel feature selection (FS) technique and in comparison
to the formula-based and Artificial Neural Network techniques, the new model has reduced Root
Mean Squared Error by around 55 and 15.The model employs the least squares support vector
machine (LS-SVM) to discover the mapping between the influencing factors and the slab
punching capacity. Furthermore, the firefly algorithm (FA), a population-based metaheuristic, is
utilized to facilitate the LS-SVM training.[16]
Nhat-DucHoang, paper, titled, “Estimating punching shear capacity of steel fibre reinforced
concrete slabs using sequential piecewise multiple linear regression and artificial neural
network”. This study uses artificial neural networks (ANN) and piecewise multiple linear
regression (PMLR) to build a prediction model that can roughly translate the mapping function
between the punching shear capacity of SFRC flat slabs and its affecting parameters. The
Levenberg-Marquardt backpropagation technique and gradient descent algorithms are used to
train the ANN-based prediction models. This data set is then used to train and verify the
sequential PMLR (SPMLR) and ANN models. Experimental results show that SPMLR can

4
deliver prediction outcome which is better than those of ANN as well as empirical design
equations.[17]
Gamze Dog˘an Musa Hakan Arslan , paper, titled, “Determination of Punching Shear
Capacity of Concrete Slabs Reinforced with FRP Bars Using Machine Learning” It stressed on
prediction models were developed for the punching strength of the slabs by using the relevant
algorithms in five different machine learning techniques (Multiple Linear Regression), Bagging-
Decision Tree Regression, Random Forest Regression, Support Vector Regression and Extreme
Gradient Boosting (MLR, Bagging-DT, RF, SVR, XGBoost).[18] The best results were achieved
by the SVR among the five different algorithms. SVR achieved a predicted success for the
strength of slabs produced with GFRP bars. After analysis, R2 values, MAE and RMSE
performance metrics were found to be well above the empirical correlations with 96.23.
Table 1. Related works.
Author Paper Title ML Model
Used
Techniques
Employed
Result References
Sujith
Mangalathua
,Hanbyeol Shin,
EunsooChoic,
Jong-Su Jeon
Explainable
machine
learning models
for punching
shear strength
estimation of
flat slabs
without
transverse
reinforcement
Linear
regression,
Ridge
regression,
support
vector
regression,
decision tree,
K-nearest
neighbors,
random
forest,
adaptive
boosting, and
extreme
gradient
boosting
Used
explainable
machine
learning
techniques like
SHAP
Extreme
gradient
boosting
model has
a
coefficient
of
determinati
on of 0.98,
and the
associated
coefficient
of variation
is 0.09.
[13]
Yuanxie Shen,
Linfeng Wu,
Shixue Liang
Explainable
machine
learning-based
model for
failure mode
identification of
RC flat slabs
without
transverse
reinforcement
Linear
Regression,
Artificial
Neural
Network,
Decision
Tree, Support
Vector
Regression,
Random
Forest,
AdaBoost,
GBDT,
XGBoost
The prediction
of XGBoost is
explained by
SHAP
XGBoost is
selected as
the best
model, in
which the
precision,
recall, F1
score and
accuracy of
which are
97.30%,
94.74%,
96.00%
and
99.02%,
respectivel
y.
[14]

5
Shasha Lu,
Mohammadreza
Koopialipoor ,
Panagiotis G.
Asteris , Maziyar
Bahri ,Danial
Jahed
Armaghani
A Novel
Feature
Selection
Approach
Based on Tree
Models for
Evaluating the
Punching Shear
Capacity of
Steel Fiber-
Reinforced
Concrete Flat
Slabs
Random
forest ,
Random tree,
and
classification
and
regression
trees (CART)
Novel feature
selection (FS)
technique has
been used.
The range
of R2 and
RMSE
values were
obtained as
0.9476–
0.9831 and
14.4965–
24.9310,
respectivel
y; in this
regard,
could be
applied to
predicting
SFRC flat
slabs.
[15]
Duy-Thang Vu ,
Nhat-Duc Hoang
Punching shear
capacity
estimation of
FRP-reinforced
concrete slabs
using a hybrid
machine
learning
approach
Least squares
support
vector
machine (LS-
SVM)
Firefly
algorithm (FA),
a population-
based
metaheuristic, is
utilised to
facilitate the
LS-SVM
training.
New model
has
achieved
roughly 55
and 15%
reductions
of Root
Mean
Squared
Error
compared
with the
Artificial
Neural
Network
methods
[16]
Nhat-Duc Hoang Estimating
punching shear
capacity of steel
fibre reinforced
concrete slabs
using sequential
piecewise
multiple linear
regression and
artificial neural
network
Piecewise
multiple
linear
regression
(PMLR)and
Artificial
neural
network
(ANN)
The algorithms
of gradient
descent and
Levenberg-
Marquardt
backpropagatio
n are employed
to train the
ANN.
[17]

6
Junho Song,
Won-Hee Kang,
Kang Su Kim,
Sungmoon Jung
Probabilistic
shear strength
models for
reinforced
concrete beams
without shear
reinforcement
Probabilistic
shear strength
models
Using a
Bayesian
Method for
parameter
estimation.
Model
predicts the
result with
improved
accuracy
and helps
incorporate
the model
uncertaintie
s into
vulnerabilit
y
estimations
and risk-
quantified
designs.
[19]
3. MATERIAL AND DATASET
The flat slab system of reinforced concrete has been used more frequently because it has some
advantages when compared to conventional structural systems[20]. Among these advantages,
one can mention greater architecture in defining internal environments or future layout changes;
simplification of reinforcement and consequent reduction of labour and material costs; ease in the
arrangement of installations and simplification of forms and framing[21]. The system also has
disadvantages compared to conventional ones, such as higher levels of vertical displacement of
the structure, reduction of the global stability and the possibility of failure by punching shear[22].
Punching shear is a type of shear failure that can occur in plate elements subjected to a
concentrated load or reaction applied transversally and is characterized by occur- ring abruptly,
which can lead the structure to ruin through progressive collapse[23]. The shear strength of the
slab-Column connection is one of the most important parameters in the design of flat slab[24].
The original file is a database created by The American Concrete Institute Committee 445C with
experimental results of 519 flat slabs tested by several authors since 1938. Experimental tests in
Civil Engineering are usually performed with reduced size structures, due to the practical issues
with testing real size structures. This is the cleaned data with fewer observations, since the goal is
to predict punching shear resistance and some of the slabs in the original dataset did not fail by
this mechanism[25].
Statistical Analysis and Data Distribution of various attributes in the dataset is as follows:-

7
Figure 1. Statistics analysis
Figure 2. Data distribution of each feature
Dataset is being divided into 87.5 % for training purposes and 12.5 % for testing purposes.
Features like Shape and d1(mm) are dropped because of its less significance and lots of null
values. To analyze each shape’s significance we converted it into numerical arrays using a one-
hot encoding method. After the analysis of Heatmap of each variable, we observed that Shape S,
C, R correlation with the target variable are significantly close to zero which indicates it is
independent of the target variable. Therefore, the rest of the 7 features will be further processed
to make a feature matrix. We have used MinMaxScaler() in Sklearn library which internally
works as following:-
In which the minimum of features is made equal to zero and the maximum of features equal to
one. MinMaxScaler() shrinks the data within the given range, usually of 0 to 1. It scales the
values to a specific value range without changing the shape of the original distribution.

8
x_std = (x-x.min (axis=0)) / x.max(axis=0) – x.min (axis=0)) (3.1)
x_scaled = x_std* (max-min) + min (3.2)
Where,
• min, max = feature_range
• x.min (axis=0) : Minimum feature value
• x.max (axis=0) : Maximum feature value
Figure 3. Null value table
Figure 4. Heatmap

9
Figure 5. One hot encoding
Figure 6. Pearson correlation coefficient
4. METHODOLOGY
4.1. Introduction to Algorithms
4.1.1. Artificial Neural Network
Figure 7. Artificial neural network
There’s huge loss in terms of financial and materials because of current civil engineering
methods. Failure during the construction phase of the project results in worker and staff fatalities.
Punching shear strength becomes exceedingly boring, as was explained earlier while employing
the old way. We can determine the shear strength by considering only a few inputs, such as the
column’s shape, size, thickness of the slab, and compressive strength of the concrete, by
employing machine learning.

10
There are a slew of traditional methods to evaluate the concrete’s compressive and tensile
strengths in order to see if it is strong enough to stand on its own. If not, determine whether the
amount of reinforcement is fair, if not create rational things. Changing the structure comprises the
following: increasing the slab’s depth, modifying the slab’s framework entails, increasing the
slab’s depth, increasing the slab’s dimensions or by utilizing reinforcement that is both vertical
and transverse.
Artificial neural network (ANN)[26] is a computing model whose layered structure resembles the
networked structure of neurons in the brain [27] . It features interconnected processing elements
called neurons that work together to produce an output function. Neural networks are made of
input and output layer/dimensions, and in most cases, they also have a hidden layer consisting of
units that transform the input into something that the output layer can use. Backpropagation
algorithm [28] is used to train the neural network.
Input x: Set the corresponding activation a1
for the input layer.
Feed forward: For each l = 2, 3,… L compute z1
= w1
a l-1
+ bl
and al
= σ(zl
). (4.1)
Output error δL
: Compute the vector δL
= ∇ a C ʘ σ'(zl
). (4.2)
Back propagate the error: For each l=L-1, L-2,…, 2 compute
δ1
= ((w l+1
)T δ1+1
)ʘ σ'(zl
). (4.3)
Output: The gradient of the cost function is given by
and (4.4)
4.1.2. Genetic Algorithm
Figure 8. Genetic algorithm

11
The genetic algorithm [29] is a heuristic search and an optimization method inspired by the
process of natural selection. It is widely used for finding a near-optimal solution to optimization
problems with large parameter space. The evolution of species (solutions in our case) is
mimicked by depending on biologically inspired components, e.g., crossover. Furthermore, as it
does not take auxiliary information into account (e.g., derivatives), it can be used for discrete and
continuous optimization.
For using a GA, two preconditions have to be fulfilled,
a) A solution representation or defining a chromosome
b) A fitness function[30]to evaluate produced solutions. In our case, a binary array is a genetic
representation of a solution (see Figure 1) and the model's Root-Mean-Square Error (RMSE) on
the validation set will act as a fitness value. Moreover, three basic operations that constitute a
GA, are as follows:
1) Selection: It defines which solutions to preserve for further reproduction e.g. roulette wheel
selection.
2) Crossover: It describes how new solutions are created from existing ones e.g. n-point
crossover.
3) Mutation: Its aim is to introduce diversity and novelty into the solution pool by means of
randomly swapping or turning-off solution bits e.g. binary mutation.
Occasionally, a technique called “Elitism” is also used, which preserves the few best solutions
from the population and passes them on to the next generation [31]. Figure 8 depicts a complete
genetic algorithm, where initial solutions (population) are randomly generated. Next, they are
evaluated according to a fitness function, and selection, crossover, and mutation are performed
afterward. This process is repeated for a defined number of iterations (called generations in GA
terminology). In the end, a solution with the highest fitness score is selected as the best solution.
4.2. Model Architecture
ANN model consists of:
1) Input layer with one neuron
2) 3 Hidden layers with 8, 6, 6 neurons
3) Output layer with one neuron
4.3. Performance Evaluation
Throughout the experiments, Root Mean Square Error (RMSE) [32] is selected to judge the
performance of the model. RMSE has 2 purposes:-
1) To serve as a heuristic for training models
2) To evaluate trained models for usefulness / accuracy RMSE is a good estimator [33] for the
standard deviation of the distribution of our errors. Formula for RMSE is as follows:

12
(4.3)
4.4. Training Approach
The training is used to assist the ANN in tuning its weights [34] in each layer to calculate a
projected output that is as near to the training label as possible. To forecast the unknown data,
GA was used to optimize the weights in each layer of the training model.
Initially, a fixed number of populations is specified. The weights of all levels in the sequential
model are generated at random for each population.
The initial phase in the GA process is population initialization. It is a subset of solutions in
current generations. There are two primary approaches for initialization of a Population in a GA:
1) Random Initialization [35]: totally random solutions are used to fill the initial population.
2) Heuristic Initialization [36]: fill the initial population using a problem-specific heuristic.
The training data will then be loaded into the training model, and the prediction process will
begin. Following the fitness calculation, which indicates how fit or good the answer is in relation
to the problem under discussion. Because it's weights are ideal, the programme will update the
maximum fitness value for the final training stage, per- haps yielding better accuracy in the final
training stage. This procedure will continue till the maximum generation is reached. The ideal
matrix will be set to ANN model after optimizing the weight matrix and will be ready to
generalize the testing data. The ANN model has one input layer, three hidden layers with 8,6,6
neurons, and one neuron in the output layer. By preventing the model from becoming locked in a
local minimum situation, a Genetic algorithm might assist improve accuracy.
The GA [37] consists of three major components:
1) selection,
2) crossover, and
3) mutation.
First, the system picks the gene pool’s elite parents. The crossover is then implemented. Among
the finest genes (weighted matrix), the process randomly picks two genes and recombines them in
the following manner:
Select a random split point for the elite genes 1 and 2. Then join the second portion of gene 2 to
the first part of gene 1, and repeat for the remaining parts of the two genes. As a result, I have
two possible elite recombined genes. Third, because mutations occur at random, they are
possible. After completing the crossover, the mechanism will create a random number between 0
and 1. If the randomly produced value is less than or equal to 0.05, a random section of the
weighted matrix will be multi- plied by another random integer between 2 and 5. By gently
scaling specific values in the weighted matrix, the mutation process can be aided in preventing
the ANN model from being trained in the wrong direction.

13
Figure 9. Algorithm
5. EXPERIMENTS
5.1 Performance Evaluation of various Training Techniques
To examine alternative training algorithms or approaches, we evaluated ANN models with SGD
(Stochastic Gradient Descent) [38], RMSprop (Root Mean Square Propagation) [39], and Adam
(Adaptive Moment Estimation) [40] training methods. The outcomes of the various algorithms
are shown below:
Table 2. Different optimizer and their corresponding RMSE score
Experimental observations
Training Method RMSE
Genetic Algorithm 0.192
SGD 0.216
RMSProp 0.176
Adam 0.160
Figure 10. Fitness v/s generation

14
Figure 11. Sgd mae vs epoch
The experiment above shows that the stochastic gradient descent approach has a difficulty with
convergence to the global minimum [41], resulting in lesser accuracy than alternative training
methods. The genetic algorithm, on the other hand, works well because it includes selection,
crossover, and mutation processes that may enhance the local minimum issue by developing
various types of genes (weighted matrix) and picking the best among them.
There are several parameters that must be hyper-tuned [42], including:
1) Number of Neurons
2) Number of generations
3) Population size
Figure 12. RMSprop mae vs epoch

15
Figure 13. Adam epoch vs mae
1) We have increased the number of neurons by changing the architecture of the neural network
defined in the code. We kept the number of layers constant.
We observed that the RMSE score lowers as the number of neurons grows, indicating that the
model is improving.
Table 3. Number of neurons and their corresponding RMSE score
Neurons RMSE
2 0.149
4 0.228
6 0.192
8 0.147
10 0.145
2) We have increased the number of generations used in the Genetic algorithm as a hyper-
parameter. We increased the no. of generations by keeping other parameters constant.
We observed that there is no specific pattern as it first in- creases then decreases.
Table 4. Number of generations and their corresponding RMSE score
Generations RMSE
10 0.181
25 0.195
50 0.184
100 0.318
3) We have increased the number of population used in Genetic algorithm as a hyper-parameter .
We increased the no. of population by keeping other parameters constant.
We observed that there is no specific pattern as it first in- creases then decreases.

16
Table 5. Number of population and their corresponding RMSE score
Populations RMSE
10 0.192
20 0.143
40 0.276
80 0.159
100 0.227
6. CONCLUSION
Civil engineering domain has emerged as one of the most significant industries all across the
countries as it helps in the development of any nation. Infrastructure building by creating
different means of transportation such as roadways, rail- ways, bridges, tunnels,
telecommunication, schools, afford- able houses etc. is a significant barometer of any country’s
development.
There are a slew of ways in which punching pressure is calculated. Getting fully interpretable and
unbiased results from these calculations is a really difficult task. A retrospective analytical
building of models shows how to calculate the punching pressure using machine learning
techniques such as ANN, genetic algorithms etc.
One major advantage of our machine learning approach is that it reduces the time period of
calculations involved.
Generally in the civil engineering process, most of the calculations are done based on hypothesis
and experiences which leads to biases and inaccuracies.
Apparently the machine learning model creates a function which takes required input and this
leads to the reduction of the time. The Gradient Descent method [43] lays the foundation for
machine learning and deep learning techniques.
The major fashion in which it has been used is a set number of populations initially specified. For
each population, the weights for every level in the sequential model are created at random.
In this study, ANN, genetic algorithms, and other optimizers are used to determine the punch
shear strength of reinforced concrete slabs.
The future efforts can be directed towards fine-tuning the neural network models using boosting
methods such as gradient boosting. The experiments here used ANN models for primary training,
but other types of deep learning models like CNN [44] and LSTM [45] can be used. Another
possible approach is to fine-tune the other hyper-parameters used in algorithms like Mutation
rate, scale value, split up point etc.
REFERENCES
[1] Richard C. Elstner and Eivind Hognestad Shearing Strength of Reinforced Concrete Slabs Journal
Proceedings Volume: 53
[2] The Maximum Punching Shear Resistance of Flat Slabs Jaroslav Halvonik Part of special
issue:CONCRETE AND CONCRETE STRUCTURES 2013 - 6th International Conference, Slovakia
[3] Effect of degree of corrosion on the properties of reinforcing steel bars Abdullah A.Almusallam
Construction and Building Materials Volume 15, Issue 8, December 2001

17
[4] Fiber-reinforced polymer composites in strengthening reinforced concrete structures: A critical
review MZ Naser, RA Hawileh, JA Abdalla - Engineering Structures, 2019 - Elsevier
[5] Fiber-reinforced polymers bars for compression reinforcement: A promising alternative to steel bars
N Elmessalami, A El Refai, F Abed - Construction and Building Materials, 2019 - Elsevier
[6]Punching of RC slabs under eccentric loads M Farzam, NA Fouad, J Grünberg… - Structural
Concrete, 2010 - icevirtuallibrary.com
[7] Application of an analytical method for the design for robustness of RC flat slab buildings P
Martinelli, M Colombo, S Ravasini, B Belletti - Engineering Structures, 2022 - Elsevier
[8] Artificial intelligence (AI) applied in civil engineering ND Lagaros, V Plevris - Applied Sciences,
2022 - mdpi.com
[9] Applications of machine learning to BIM: A systematic literature review A Zabin, VA González, Y
Zou, R Amor - Advanced Engineering Informatics, 2022 - Elsevier
[10] Damage assessment of RC flat slabs partially collapsed due to punching shear T Cosgun, B Sayin -
International Journal of Civil Engineering, 2018 - Springer
[11] Ultimate punching shear strength analysis of slab–column connections DD Theodorakopoulos, RN
Swamy - Cement and Concrete Composites, 2002 - Elsevier
[12] Prediction of punching shear capacity of RC flat slabs using artificial neural network NA Safiee, A
Ashour - Asian Journal of Civil Engineering, 2017 - sid.ir
[13] Mangalathu, Sujith, Hanbyeol Shin, Eunsoo Choi, & Jong-Su Jeon, (2021) "Explainable machine
learning models for punching shear strength estimation of flat slabs without transverse
reinforcement", Journal of Building Engineering, Vol. 39, pp 102300.
[14] Shen, Yuanxie, Linfeng Wu, & Shixue Liang, (2022) "Explainable machine learning-based model for
failure mode identification of RC flat slabs without transverse reinforcement", Engineering Failure
Analysis,Vol. 141, pp 106647.
[15] Lu, Shasha, Mohammadreza Koopialipoor, Panagiotis G. Asteris, Maziyar Bahri, & Danial Jahed
Armaghani, (2020) "A novel feature selection approach based on tree models for evaluating the
punching shear capacity of steel fiber-reinforced concrete flat slabs", Materials, Vol. 13, pp 173902.
[16] Vu, Duy-Thang, & Nhat-Duc Hoang, (2016) "Punching shear capacity estimation of FRP-reinforced
concrete slabs using a hybrid machine learning approach", Structure and Infrastructure
Engineering,Vol. 12, no. 9, pp 1153-1161.
[17] Hoang, Nhat-Duc, (2019) "Estimating punching shear capacity of steel fibre reinforced concrete slabs
using sequential piecewise multiple linear regression and artificial neural network",
Measurement,Vol. 137, pp 58-70.
[18] Doğan, Gamze, & Musa Hakan Arslan, (2022) "Determination of Punching Shear Capacity of
Concrete Slabs Reinforced with FRP Bars Using Machine Learning", Arabian Journal for Science
and Engineering, pp 1-27.
[19] Song, Junho, Won-Hee Kang, Kang Su Kim, & Sungmoon Jung, (2010) "Probabilistic shear strength
models for reinforced concrete beams without shear reinforcement", Structural engineering &
mechanics, Vol. 11, no. 1, pp 15.
[20] Comparative Analysis of Diagrid Structural System and conventional structural system for high
rise steel building H Varsani, N Pokar, D Gandhi - International Journal of Advance …, 2015 -
academia.edu
[21] Simplified diverse embedment model for steel fiber-reinforced concrete elements in tension SC Lee,
JY Cho, FJ Vecchio - Materials Journal, 2013 - vectoranalysisgroup.com
[22] Analysis of precision of geodetic instruments for investigating vertical displacement of structures B
Kovačič, A Pustovgar, N Vatin - Procedia engineering, 2016 - Elsevier
[23] Dynamic stability of suddenly loaded structures GJ Simitses - 2012 - books.google.com
[24] Ultimate punching shear strength analysis of slab–column connections DD Theodorakopoulos, RN
Swamy - Cement and Concrete Composites, 2002 - Elsevier
[25] Prediction of punching shear capacity for fiber-reinforced concrete slabs using neuro-nomographs
constructed by machine learning E Alotaibi, O Mostafa, N Nassif, M Omar… - Journal of Structural
…, 2021 - ascelibrary.org
[26] A general regression neural network DF Specht - IEEE transactions on neural networks, 1991 -
academia.edu
[27] How neural networks learn from experience GE Hinton - Scientific American, 1992 - JSTOR
[28] The backpropagation algorithm R Rojas - Neural networks, 1996 - Springer
[29] Genetic algorithms S Mirjalili - Evolutionary algorithms and neural networks, 2019 - Springer

18
[30] A genetic algorithm fitness function for mutation testing L Bottaci - Proceedings of the SEMINALL-
workshop at the 23rd …, 2001 - Citeseer
[31] Genetic algorithm performance with different selection strategies in solving TSP NM Razali, J
Geraghty - Proceedings of the world congress on …, 2011 - iaeng.org
[32] Root mean square error (RMSE) or mean absolute error (MAE)?–Arguments against avoiding RMSE
in the literature T Chai, RR Draxler - Geoscientific model development, 2014 - gmd.copernicus.org
[33] Advantages of the mean absolute error (MAE) over the root mean square error (RMSE) in assessing
average model performance
[34] An efficient optimization approach for designing machine learning models based on genetic
algorithm KM Hamdia, X Zhuang, T Rabczuk - Neural Computing and Applications, 2021 - Springer
[35] Gradient descent with random initialization: Fast global convergence for nonconvex phase retrieval Y
Chen, Y Chi, J Fan, C Ma - Mathematical Programming, 2019 - Springer
[36] Evolutionary heuristic a* search: Heuristic function optimization via genetic algorithm YF Yiu, J Du,
R Mahapatra - 2018 IEEE First International …, 2018 - ieeexplore.ieee.org
[37] Genetic algorithm-A literature review A Lambora, K Gupta, K Chopra - 2019 international
conference …, 2019 - ieeexplore.ieee.org
[38] Is local SGD better than minibatch SGD? B Woodworth, KK Patel, S Stich, Z Dai… - International
…, 2020 - proceedings.mlr.press
[39] Improved adam optimizer for deep neural networks Z Zhang - 2018 IEEE/ACM 26th International
Symposium on …, 2018 - ieeexplore.ieee.org
[40] Convergence of the RMSProp deep learning method with penalty for nonconvex optimization
D Xu, S Zhang, H Zhang, DP Mandic - Neural Networks, 2021 - Elsevier
[41] Gradient descent finds global minima of deep neural networks S Du, J Lee, H Li, L Wang… - …
conference on machine …, 2019 - proceedings.mlr.press
[42] Extending MLP ANN hyper-parameters Optimization by using Genetic Algorithm F Itano, MAA de
Sousa… - 2018 International joint …, 2018 - ieeexplore.ieee.org
[43] Backpropagation and stochastic gradient descent method S Amari- Neurocomputing, 1993 - Elsevier
[44] A CNN regression approach for real-time 2D/3D registration S Miao, ZJ Wang, R Liao - IEEE
transactions on medical …, 2016 - ieeexplore.ieee.org
LIST OF ABBREVIATIONS
ANN - Artificial neural network
EA - Evolutionary algorithms
SGD - Stochastic gradient descent
FRP - Fiber reinforced polymer
LR - Linear Regression
DT - Decision Tree
SVC - Support vector clustering
RF - Random Forest
GBDT - Gradient Boosting Decision Tree
XGBoost - eXtreme Gradient Boosting
SHAP - SHapley Additive exPlanations
random tree - (RT)
(CART) - classification and regression trees
firefly algorithm - (FA),
feature selection - (FS)
least squares support vector machine - (LS-SVM)
piecewise multiple linear regression - (PMLR)
sequential PMLR - (SPMLR)
MLR - Multiple Linear Regression
RMSE - Root-Mean-Square Error
R2 score - r squared score
RMSprop - Root Mean Square Propagation

19
Adam - Adaptive Moment Estimation
Mae - Mean absolute error
CNN - Convolutional neural network
LSTM - Long short-term memory
AUTHORS
Gaurav Sarkar
Gaurav Sarkar is a final year undergraduate student of the computer science branch. He
is skilled in machine learning and deep learning. Experienced in Designing and
developing ML Models and systems across diverse industries. Also proficient in
collaborating with teams of high performing professionals and won several medals on
kaggle platform and hackathon.
Suhail Ahmad Magray
Sohail Ahmed Margray is a Final year undergraduate student of the civil engineering
branch. He is creative while solving problems Adept with civil engineering tools And
creates 2D drawings and designs using AutoCAD. He is proficient in collaborating with
individuals of various backgrounds. Also Designs concrete structural elements, e.g.
foundation, beams and walls.
Tanushree Sarkar
Tanushree Sarkar is a Ph.D. scholar of biotechnology department. She is skilled in
working with development of bioproducts and bioprocessing. Also experienced in the
field of bioinformatics.

Prediction of Punching Shear Strength using Metaheuristic Approach of Optimization

More Related Content

Similar to Prediction of Punching Shear Strength using Metaheuristic Approach of Optimization (20)

More from ADEIJ Journal (20)

Recently uploaded (20)

Prediction of Punching Shear Strength using Metaheuristic Approach of Optimization